An environment vector map automatic construction method for unmanned vehicle navigation

Abstract

The application discloses an environment vector map automatic construction method for unmanned vehicle navigation, and can automatically generate an unmanned lane-level high-precision map, and comprises the following points: the method mainly adopts three kinds of sensor data of a laser radar, an IMU and an RTK-GPS, in an environment where satellite signals are refused in a tunnel and the like, laser positioning results are adopted, in an open environment, satellite signals are better, and RTK-GPS satellite navigation positioning results and laser positioning can be used.

Description

Technical Field

[0001] This invention belongs to the field of high-precision map technology for autonomous driving, specifically a method for automatically constructing environmental vector maps for autonomous vehicle navigation. Background Technology

[0002] Perception is a crucial prerequisite for autonomous driving. Low-level autonomous driving relies primarily on sensor data, which is insufficient in adverse weather conditions or when obstructed, failing to guarantee the safe and autonomous driving of autonomous vehicles. Lane-level high-precision maps can provide autonomous vehicles with prior path and navigation decision-making information, and have become an essential condition for high-level autonomous driving, serving as the main data source for the perception module of autonomous vehicles.

[0003] Environmental vector maps are a type of lane-level high-precision map. They can describe the relationships between lanes and various traffic elements in detail. Road-level and lane-level road networks are the core components of environmental vector maps. Currently, Level 4 autonomous driving is still in the research stage both domestically and internationally, and the lack of lane-level high-precision maps is a significant factor limiting the development of advanced autonomous driving technologies. Regarding high-precision map models for autonomous driving, there is currently no complete and universally applicable map model; development is moving towards layered and highly detailed models. In terms of creating high-precision maps for autonomous driving, current methods mainly include lane-level map creation based on recorded driving trajectories and software-based manual annotation methods. Lane-level map creation based on recorded driving trajectories forms a road network, but the traffic network at intersections is fixed during recording, resulting in inflexible and incomplete intersection traffic networks. Furthermore, it requires recording multiple lane trajectories, leading to a large data collection workload. Manual annotation methods have low accuracy and are not suitable for large-scale map annotation. Summary of the Invention

[0004] To address the shortcomings of methods that involve cumbersome operation and inability to construct complete road networks when recording driving trajectory maps, and the low accuracy and unsuitability of manual map annotation methods for large-scale map annotation, this invention proposes an automatic environmental vector map construction method for autonomous vehicle navigation, further improving the accuracy of lane-level high-precision autonomous driving.

[0005] This invention is achieved through the following technical solution:

[0006] Includes the following steps:

[0007] Step 1: Acquire raw data from three sensors: lidar, inertial measurement unit, and RTK-GPS satellite positioning / laser positioning, and perform data preprocessing;

[0008] Step 2: Using the collected data from the lidar and inertial measurement unit, an obstacle point cloud map is created using SC-LEGO-LOAM;

[0009] Step 3: Using the lateral lane expansion method based on high-precision positioning, the road centerline, road boundary line, lane centerline, lane boundary line, and entry and exit nodes of the road segment are obtained;

[0010] Step 4: Use the DBSCAN intersection clustering algorithm to cluster the entry and exit nodes of the road segment, so that the entry and exit nodes that belong to the same intersection in space belong to the same intersection cluster, and then construct the intersection traffic matrix.

[0011] Step 5: Generate the intersection driving trajectory using the intersection trajectory generation method based on Dubins curves within the intersection;

[0012] Step 6: Construct the road network at the road layer and the lane network at the lane layer, and save the environmental vector map.

[0013] 1. In step one, RTK-GPS obtains latitude, longitude, and altitude data in the WGS-84 coordinate system. Data preprocessing requires unifying RTK-GPS satellite positioning and NDT laser positioning into the same rectangular coordinate system. The steps to convert RTK-GPS to a rectangular coordinate system include the following four steps:

[0014] S1: Convert the positioning results (lon,lat,alt) in the WGS-84 coordinate system to ECEF coordinates (x... ECEF ,y ECEF ,z ECEF The conversion relationship is shown in Equation 1, where a is the Earth's radius and e is the eccentricity.

[0015]

[0016] S2: Set the origin of the ECEF coordinate system. Set the initial ENU coordinates as the origin (0, 0, 0) in the ENU coordinate system, and the corresponding ECEF coordinates as the reference coordinate point of the ECEF.

[0017] O ECEF =(x0,y0,z0) (2)

[0018] S3: Any point P in space ECEF = (x,y,z) corresponds to ENU rectangular coordinates P ENU P obtained ENU The coordinates serve as the high-precision positioning information for step two.

[0019]

[0020] This is the pose transformation matrix from the ECEF coordinate system to the ENU coordinate system, determined by the vehicle's initial attitude angles roll, pitch, and yaw. The calculation relationship is as follows:

[0021]

[0022] 2. In step three, the specific process of the lateral lane expansion method based on high-precision positioning includes:

[0023] S1: Acquire high-precision positioning data of the road centerline or the centerline of a lane. The data acquisition vehicle travels along the road centerline or the centerline of any lane, receiving RTK-GPS or NDT laser positioning data to obtain the spatial location information of the road or lane centerline.

[0024] S2: Calculate reference point C, located to the left of the data collection vehicle and perpendicular to the road's direction of travel, at a distance of one lane width from the vehicle's trajectory. A and B are two points relatively close to each other on the trajectory, and O is located between A and B. Then reference point C satisfies Equation 5:

[0025]

[0026] S3: Calculate all lane shape control points on the transverse direction of the road (straight line CO). On a standard road, lanes in the same direction of travel are parallel to each other and spaced equally. Therefore, given points C and O, all lane control points in the CO direction can be calculated using the principle of a constant ratio.

[0027] 3. In step four, the DBSCAN intersection clustering algorithm is used to cluster the entry and exit nodes of the road segment. This specifically includes:

[0028] S1: Set the algorithm parameters and use the DBSCAN algorithm to cluster all the entering and exiting nodes in the space to form an intersection cluster.

[0029] S2: Establish a mathematical model of the intersection cluster, construct a road-level intersection traffic matrix and a lane-level intersection traffic matrix, and use the traffic matrix to describe the traffic relationship between each entering and exiting node within the intersection.

[0030] 4. In step five, the intersection driving trajectory generated by the Dubins curve-based intersection driving trajectory generation method meets the minimum turning radius and heading requirements for vehicles entering and exiting nodes. The specific steps are as follows:

[0031] S1: Extract the entry and exit nodes of the intersection. The entry and exit nodes include spatial location coordinates and heading, respectively s(x i ,y i ,α i ), g(x g ,y g ,βg The following steps, S2-S5, are used to generate the intersection driving trajectory.

[0032] S2: Perform coordinate transformation on the entry and exit nodes, with the new coordinates being s(0,0,α) and g(d,0,β).

[0033]

[0034] S3: Calculate the θ, t, p, and q parameters for the four types of curves: LSL, RSR, LSR, and RSL. The calculation methods are shown in formulas 7, 8, 9, and 10 respectively.

[0035]

[0036]

[0037]

[0038]

[0039] S4: Calculate the total length of the four types of curves: length = t + p + q, and select the curve type with the smallest length as the curve type Γ of the target trajectory;

[0040] S5: Traverse the points i on Γ, and push the path points into the path sequence segment by segment. The pose calculation of the trajectory points of segments L, R, and S is shown in formulas 11, 12, and 13, respectively:

[0041] x+=sin(phi+v)-sinphi,y+=-cos(phi+v)+cosphi,yaw=phi+v (11)

[0042] x+=-sin(phi-v)+sinphi,y+=cos(phi-v)-cosphi,yaw=phi-v (12)

[0043] x+=vcosphi, y+=vsinphi, yaw=phi (13)

[0044] Restore regularity to x and y: x = x * r + x i y=y*r+y i ;

[0045] Add (x,y,yaw) to the path point sequence path to form the point sequence of the driving trajectory at the intersection. Attached Figure Description

[0046] Figure 1The flowchart for the automatic construction of environmental vector maps for unmanned vehicle navigation according to the present invention mainly includes the steps of data acquisition, lateral lane expansion, and intersection topology traffic relationship construction.

[0047] Figure 2 This is a schematic diagram illustrating the calculation of reference point C on the left side of the data acquisition vehicle in step three of this invention.

[0048] Figure 3 This diagram illustrates the calculation of lane control points using the lateral expansion method in step three of this invention, serving to more intuitively and clearly explain the principle of lateral lane expansion. A and B are two points on the trajectory of the data acquisition vehicle, and C is a reference point on the left side of the vehicle. The lateral lane expansion method can then be used to calculate control points D for lane 2, E for lane 3, and F for lane 4. By continuously collecting data along the road segment, an ordered sequence of shape control points for all lanes on the road segment can be obtained.

[0049] Figure 4 The obstacle point cloud map result is an example of the present invention, with a resolution of 0.6m;

[0050] Figure 5 This is a road network layer map result of an example of the present invention, which consists of the road center of each road segment. The traffic relationship network at intersections is not shown in the figure.

[0051] Figure 6 This is a road network layer map result of an example of the present invention, which is a road-level topology network composed of road centerlines and intersection traffic relationships of road segments;

[0052] Figure 7 The lane network layer map result is an example of the present invention, which is a lane-level topology network composed of the lane center lines of road segments and the traffic relationships between lanes within intersections.

[0053] Beneficial effects

[0054] This invention provides a method for producing lane-level high-precision maps for autonomous vehicles—an automatic environmental vector map construction method for autonomous vehicle navigation. This method only requires collecting centimeter-level high-precision positioning data of one road centerline or one lane centerline to automatically construct all lane centerlines and lane boundary lines on the road. Based on the entry and exit points of lanes, it automatically constructs a road intersection network, thereby building a road-level and lane-level road network for autonomous vehicle navigation, guiding autonomous vehicles to drive automatically. Compared to current methods of constructing high-precision maps by recording driving trajectories and software-based manual annotation, this automatic environmental vector map construction method for autonomous vehicle navigation does not require collecting data from multiple lanes. It can construct flexible intersection networks at intersections, is simpler to collect data than methods that record driving trajectories, and constructs a more flexible and complete road network. Compared to software-based manual annotation methods, it offers higher accuracy and is suitable for large-scale map construction. Detailed Implementation

[0055] The following examples provide a more detailed and complete description of the automatic environmental vector map construction method for unmanned vehicle navigation provided by this invention. The examples described below are exemplary and are only used to explain this invention, and should not be construed as limiting the invention.

[0056] Step 1: Acquire raw data from three sensors: LiDAR, IMU, and RTK-GPS, and perform data preprocessing. Use coordinate transformation to convert the satellite positioning results into Cartesian coordinates.

[0057] Step 2: Using the collected data from the lidar and inertial measurement unit, an obstacle point cloud map is created using SC-LEGO-LOAM;

[0058] Step 3: Using the lateral lane expansion method based on high-precision positioning, obtain all road centerlines, road boundary lines, lane centerlines, lane boundary lines, and entry and exit nodes of the road segment;

[0059] S1: Acquire high-precision positioning data of the road centerline or lane centerline. The data collection vehicle travels along the road centerline or the centerline of any lane, receiving RTK-GPS or NDT laser positioning data to obtain the spatial location information of the road or lane centerline.

[0060] S2: Calculate reference point C, which is located to the left of the data collection vehicle and perpendicular to the road traffic direction;

[0061] S3: Calculate all lane shape control points on the transverse side of the road (straight line CO). On standard roads, lane spacing is the same in the same direction of traffic. Therefore, given points C and O, all lane control points in the CO direction can be calculated using the principle of a constant ratio.

[0062] Step 4: Use the DBSCAN intersection clustering algorithm to cluster the entering and exiting nodes, so that nodes that belong to the same intersection in space belong to the same intersection cluster, and then construct the intersection traffic matrix.

[0063] S1: Set parameters for the entry and exit nodes, and use the DBSCAN algorithm to cluster all entry and exit nodes in the space to form an intersection cluster.

[0064] S2: Establish a mathematical model for the intersection cluster, construct the road-level intersection traffic matrix and the lane-level intersection traffic matrix, as follows:

[0065] Intersection clustering algorithm based on DBSCAN principle:

[0066]

[0067]

[0068] Step 5: Generate the intersection driving trajectory using a Dubins curve-based intersection trajectory generation method. The algorithm is as follows:

[0069] Intersection driving trajectory generation algorithm based on Dubins curve:

[0070]

[0071]

[0072] Step 6: Construct the road network at the road layer and the road network at the lane layer, and save the vector map for use in autonomous vehicle navigation planning.

[0073] Validation Example

[0074] The experiment was conducted in an urban environment near the Changqing District People's Government in Jinan City, Shandong Province. The experimental area was approximately 2km x 2km, encompassing 41 road segments and 17 intersections. Road segment types included two-lane, four-lane, and six-lane roads in both directions, while intersection types included T-junctions and crossroads. Data was collected from vehicles in this environment on February 17, 2023. Vehicles traveled along the centerline of the leftmost lane of each road segment, with an average speed of 30-40 km / h throughout the journey. Data was collected using LiDAR, IMU, and RTK-GPS to create an environmental vector map of the urban environment. The results are as follows: Figure 4 -- Figure 7 As shown.

[0075] Figure 4 It is a constructed obstacle point cloud map with a resolution of 0.6m, used to describe the obstacle features on both sides of the road, and also serves as a global map for laser positioning and matching.

[0076] Figure 5 and Figure 6 This is a constructed road network layer map used for road-level path planning and road-level path extraction. This layer map consists of road centerlines of road segments and intersection topological traffic networks. The road segments are ordered shape control points generated using a high-precision positioning-based lateral lane expansion method, with their order consistent with the road's traffic direction. The intersection topological traffic network is described by the intersection traffic matrix and driving trajectories.

[0077] Figure 7 This is a constructed lane network layer map used for lane-level path planning and lane-level path extraction. This layer map consists of lane centerlines for road segments and intersection topological traffic networks. The lane centerlines are ordered shape control points generated using a high-precision positioning-based lateral lane expansion method, with their order consistent with the road's travel direction. The intersection topological traffic network is described by the intersection traffic matrix and driving trajectories. Due to the complexity of the drivable trajectories at intersections in the lane network layer, therefore... Figure 7 Not displayed.

[0078] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically constructing environmental vector maps for unmanned vehicle navigation, characterized in that, Includes the following steps: Step 1: Acquire raw data from three sensors: lidar, inertial measurement unit, and RTK-GPS satellite positioning / laser positioning, and perform data preprocessing; Step 2: Using the collected data from the lidar and inertial measurement unit, an obstacle point cloud map is created using SC-LEGO-LOAM. Step 3: Using the lateral lane expansion method based on high-precision positioning, the road centerline, road boundary line, lane centerline, lane boundary line, and entry and exit nodes of the road segment are obtained; Step 4: Use the DBSCAN intersection clustering algorithm to cluster the entry and exit nodes of the road segment, so that the entry and exit nodes that belong to the same intersection in space belong to the same intersection cluster, and then construct the intersection traffic matrix. Step 5: Generate the intersection driving trajectory using the Dubins curve-based intersection trajectory generation method. The trajectory generated by the Dubins curve-based intersection driving trajectory generation algorithm satisfies the minimum turning radius and heading requirements of the vehicle, and has the shortest path, consisting of three segments: left turn (L), right turn (R), and straight ahead (S). Specific steps include: S1: Extract the entry and exit nodes of the intersection. The entry and exit nodes include spatial location coordinates and headings, respectively. , The following steps S2-S5 are used to generate the intersection driving trajectory; S2: Perform coordinate transformation on the entry and exit nodes; the new coordinates are... , ; ； S3: Calculate the values ​​of the four types of curves: LSL, RSR, LSR, and RSL respectively. The parameters are as follows: ； ； ； ； S4: Calculate the total length of the four types of curves. Select the curve type with the smallest length as the curve type of the target trajectory. ; S5: Traversal Points on The path points are pushed into the path sequence in segments. The trajectory pose calculation formulas for segments L, R, and S are as follows: ； ； ； right Restoring regularization: ; Will Add path point sequence This constitutes a sequence of points that form the driving trajectory at the intersection; Step 6: Construct the road network at the road level and the lane network at the lane level, and save the environmental vector map.

2. The method for automatically constructing environmental vector maps for unmanned vehicle navigation as described in claim 1, characterized in that, In step one, data preprocessing requires unifying the RTK-GPS satellite positioning results and NDT laser positioning results into the same Cartesian coordinate system. The steps to convert RTK-GPS to a Cartesian coordinate system include the following four steps: S1: Positioning results in the WGS-84 coordinate system Convert to ECEF coordinates The transformation relationship is as follows: ； in For the Earth's radius, Eccentricity; S2: Set the origin of the ECEF coordinate system. Set the initial ENU coordinates as the origin (0, 0, 0) in the ENU coordinate system, and the corresponding ECEF coordinates as the reference coordinate point of the ECEF. ； S3: Any point in space The corresponding ENU rectangular coordinates are , obtained Coordinates are used as the high-precision positioning input for step two: ； It is the pose transformation matrix from the ECEF coordinate system to the ENU coordinate system, which is derived from the vehicle's attitude angles at the initial moment. The decision and calculation relationships are as follows: 。 3. The method for automatically constructing environmental vector maps for unmanned vehicle navigation as described in claim 1, characterized in that, In step three, the specific process of the lateral expansion method based on high-precision positioning includes: S1: Obtain high-precision positioning data of the road centerline or the centerline of a lane. Allow the data collection vehicle to drive along the road centerline or the centerline of any lane and receive RTK-GPS or NDT laser positioning to obtain the spatial location information of the road or lane centerline. S2: Calculate a reference point C located to the left of the data collection vehicle and perpendicular to the road traffic direction, at a distance of one lane width from the vehicle's trajectory. A and B are two points close to each other on the trajectory, and O is located between A and B. Then, reference point C satisfies the following: ； S3: Calculate all lane shape control points on the transverse CO line of the road; on a standard road, lanes in the same direction of traffic are parallel to each other and have the same spacing. Therefore, given points C and O, all lane control points in the CO direction can be calculated based on the principle of a fixed ratio.

4. The method for automatically constructing environmental vector maps for unmanned vehicle navigation as described in claim 1, characterized in that, Step four, which uses the DBSCAN intersection clustering algorithm to cluster the entering and exiting nodes, specifically includes: S1: Set the algorithm parameters, set the minimum sample number threshold to 2, and set the neighborhood parameters according to the size of the intersection space. Use the DBSCAN algorithm to cluster all the entering and exiting nodes in the space to form an intersection cluster. S2: Establish a mathematical model of the intersection cluster, construct a road-level intersection traffic matrix and a lane-level intersection traffic matrix, and use the traffic matrix to describe the traffic relationship between each entering and exiting node within the intersection.

Images